tongjingqi

tongjingqi / AI-Can-Learn-Scientific-Taste

Public

We propose Reinforcement Learning from Community Feedback (RLCF), a training paradigm that uses large-scale community signals as supervision, and formulate scientific taste learning as a preference modeling and alignment problem.

tongjingqi.github.ioAI-Can-Learn-Scientific-Taste agent ai-innovator ai-scientists rl
44
0
100% credibility
Found Mar 17, 2026 at 44 stars -- GitGems finds repos before they trend. Get early access to the next one.
Sign Up Free
AI Analysis
AI Summary

This repository is a landing page for a research project demonstrating that AI can learn 'scientific taste' to judge paper impact and generate promising research ideas, linking to an arXiv paper, Hugging Face models/datasets, and a project website.

How It Works

1
🔍 Discover the project

You come across 'AI Can Learn Scientific Taste,' a cool research idea where AI picks up what makes science great.

2
📖 Read the introduction

You learn how AI trains on scientists' real choices from paper citations to build good judgment.

3
💡 Meet the AI helpers

Get thrilled by Scientific Judge that spots promising papers and Scientific Thinker that sparks high-impact ideas.

4
🌐 Visit project homes

Check out the colorful project page and Hugging Face spot for more details, pictures, and ready-to-use goodies.

5
⚖️ Compare research papers

Hand it two paper ideas and it tells you which one scientists would likely love more.

6
🧩 Dream up new ideas

Give it one paper and it suggests fresh follow-up ideas with big potential.

🎉 Supercharge your science

Now you have an AI buddy helping judge and create top-notch research every day.

Sign up to see the full architecture

5 more

Sign Up Free

Star Growth

See how this repo grew from 44 to 44 stars Sign Up Free
Repurpose This Repo

Repurpose is a Pro feature

Generate ready-to-use prompts for X threads, LinkedIn posts, blog posts, YouTube scripts, and more -- with full repo context baked in.

Unlock Repurpose
AI-Generated Review

What is AI-Can-Learn-Scientific-Taste?

This project proposes Reinforcement Learning from Community Feedback (RLCF), a paradigm that uses large-scale community signals like citations to train AI models on scientific taste—judging paper impact and proposing high-potential research ideas. It delivers two Hugging Face models: Scientific Judge, which compares paper pairs to predict higher-impact ones, and Scientific Thinker, which generates follow-up ideas aligned with community preferences. Built on SciJudgeBench from 2.1M arXiv papers, it formulates learning as preference modeling and alignment, with no code here—just model access and a project page.

Why is it gaining traction?

It stands out by turning citation data into actionable feedback for RL training, letting AI learn to propose ideas that match real-world impact, unlike generic ideation tools. Developers notice the generalization across fields and time, plus strong results beating baselines in pairwise comparisons. The hook is plug-and-play HF models for research workflows, echoing how you'd github propose changes in pr but for scientific alignment.

Who should use this?

AI researchers evaluating paper quality or brainstorming follow-ups in CS, math, or physics. ML engineers building reward models for open-ended generation. Academics needing quick impact prediction on arXiv preprints, like proposing day code github for next experiments.

Verdict

Grab the HF models if you're in research ideation—they work out of the box for preference-based tasks. But with 44 stars, 1.0% credibility score, and just READMEs (no tests or code), it's raw research; propose changes via issues if you want production maturity.

(178 words)

Sign up to read the full AI review Sign Up Free

Similar repos coming soon.